Discussions
Solve problems, and share tips and tricks with other JMP users.
Hello @KernDawg, The convergence criteria and optimization algorithm for Fit Mixed and Fit Least Squares (with a random effect specified) are going to be different; which I suspect is also part of the problem here. Another issue with this dataset is the covariance structure (high confounding risk), which can cause convergence problems in my experience:
Note that as a rule of thumb, we might consider a Pearson correlation of ~ 0.3 to indicate moderate risk, with ~0.5 or higher to indicate higher risk (but this will depend on subject-matter-knowledge and your own personal risk-definition).
The Fit Mixed platform may be more sensitive to this confounding risk and hence, what appears to be a convergence problem (where effect estimates are not reported). Agree with @Phil_Kay that low signal to noise ratio in this model may contribute to the problem we are seeing.
The Fit Least Squares personality with a random effect specified looks like a perfectly acceptable model specification for your modeling scenario here; in which case, none of your effects are statistically significant on your Yield Actual response (except for the Intercept term).
The fact that only the intercept is "significant" is likely another clue. You might want to look at a better-fitting model here, or consider why your expected model is not fitting your data particularly well (Adj-R-squared = 6% and is less than 1/2 of R-squared = 13%):
@KernDawg: I'd like to add one more thing here as food for thought. If we pay attention to the data structure in the "S Rate" and the "N Rate" variables, we can see that they are both rather discrete (not strictly continuous), even if they are modeled in your data table this way. You can see this in the pattern of the data as shown in the normal quantile plots. This type of pattern is often indicative of limited measurement resolution (in the context of a measurement system used to collect data).
If we fit an exponential distribution to these data (using the Distribution Platform and selecting this distribution fit for the purposes of demonstration and because Johnson Sb risks overfitting to the data), we can ask JMP to save a simulated continuous column for these fits for each of these two variables:
When we do this, and we try fitting the same model in Fit Mixed (using these variables instead of the original ones), the model converges and we successfully obtain Fixed Effects Estimates:
